{"id":"f3b59865-a012-4930-a565-a6ce26f9dd0f","arxiv_id":"2608.06376","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"By explicitly adding discrete topological sectors to a Gaussian neural sampler, the paper reproduces BKT vortex physics, string T-duality, and an exact compact rotor, demonstrating a template for compact NN-FT.","lead":"This paper shows how a neural network field theory can handle compact spaces: use a random Fourier feature sampler for smooth fluctuations and add discrete labels for topological sectors like vortices, momentum, and winding. It applies this recipe to the BKT transition and bosonic string T-duality, reproducing known physics, and adds an exact compact rotor check.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The BKT spin-wave scaling in Eq. 3.5 depends on an unspecified amplitude/cutoff normalization; with fixed A, p(n)∝|n|^{-2} gives η = b^2 A^2/(2 log Λ), not η = b^2.","rationale":"The reader's weakest assumption identifies the same issue I find, and my independent read of Eq. (3.3) confirms it. The crucial point is that the spectral density p(n)∝|n|^{-2} has divergent normalization on Z^2, so every finite-L implementation must introduce a cutoff Λ, and A must be tied to Λ for the claimed exponent to hold. Eq. (3.5) is stated without derivation or parameter specification. This is load-bearing because the entire BKT reproduction—the Gaussian critical line, b_c, T_c, and the Nelson–Kosterlitz jump—flows from that exponent; if the exponent is wrong, the apparent transition is an artifact of the arbitrary amplitude. The exact rotor check and the T-duality sector are unaffected, so I do not reject the paper; I agree with the CONDITIONAL verdict and recommend no change. A single numerical re-derivation of Eq. (3.5) with explicit A and Λ would resolve the issue, and since no code is shipped, this cannot currently be checked from the manuscript.","tokens_in":7207,"tokens_out":11157,"duration_ms":106812,"concrete_test":"Implement the spin-wave sampler (3.3) on an L=128 torus with modes truncated at Λ, draw N=10^4 frequencies from p(n)∝|n|^{-2}, fix A=1, and measure ⟨exp(ib(θ_sw(x)−θ_sw(0)))⟩ for b∈{0.25,0.5,0.75} by Monte Carlo over phases and frequencies; extract η(b) from log-log fits. If η(b)/b^2 is not 1, Eq. (3.5) fails at fixed A. Then repeat with A=√(2 log Λ) and check that η(b)=b^2 emerges. Either outcome settles whether the critical line and b_c=1/2 are predicted by the sampler or inserted through the amplitude.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The BKT demonstration is the load-bearing part of the central claim, and it rests on Eq. (3.5): θ_sw in (3.3) is asserted to give ⟨V_b(x)V_b^*(0)⟩ ∼ |x|^{-b^2}. For the Gaussian field (3.3), this correlator is exp[-(b^2/2)Var(θ_sw(x)-θ_sw(0))], and Var is set by the sampled covariance. With frequencies drawn from p(n)∝|n|^{-2} on an L×L torus with mode cutoff Λ, the dimensionless average satisfies 1−E_p cos(k·x) ≈ log|x|/log Λ at large separation, so Var ≈ A^2 log|x|/log Λ. The resulting spin-wave anomalous dimension is η_sw = b^2 A^2/(2 log Λ), not b^2. Equality η_sw = b^2 holds only if A = √(2 log Λ), or an equivalent normalization of the spectral density is imposed. The paper never specifies A or the cutoff; moreover p(n)∝|n|^{-2} is not normalizable on Z^2 without a regularization. Therefore Eq. (3.5) is an unstated matching condition, and the entries b_c = 1/2, K_0 = 1/(2πb^2), and T_c = π/2 in (3.6) are inputs imposed to reproduce standard BKT values rather than outputs of the neural sampler. Because the vortex correlator (3.9) multiplies G_sw, this free normalization shifts the entire apparent critical line. The rotor section (Sec. 5) and the T-duality lattice identity (4.6) are exact and do not share this flaw, so the concern is localized but central to the BKT claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a mixed continuous/discrete latent-variable extension of neural network field theory (NN-FT) for compact theories. The construction pairs a local Gaussian random-feature sampler with explicit sums over topological sectors: vortex sectors for the BKT transition, momentum/winding sectors for the bosonic string, and winding sectors for the quantum rotor. The BKT section claims to reproduce the Gaussian critical line, vortex proliferation, the essential singularity, and the Nelson–Kosterlitz jump; the string section claims to reproduce circle T-duality, Buscher rules, self-dual current algebra, and a toy T-fold; the rotor section gives an exact analytic check against the thermal quantum rotor. The abstract frames the common lesson as the same local sampler plus different discrete topological data producing physically distinct compact theories.","tokens_in":7593,"tokens_out":3502,"duration_ms":38269,"significance":"If the BKT demonstration were fully established, the paper would provide a concrete and conceptually useful template for extending NN-FT beyond local Gaussian physics to theories with global structure, and the rotor section in particular is a clean exact benchmark. The string section is largely kinematical and rests on standard exact lattice identities. However, the central BKT claim currently relies on an unstated normalization condition in Eq. (3.5), which shifts the critical coupling and undermines the claim that the sampler reproduces BKT rather than being tuned to it. The rotor computation is exact and machine-checkable, and the T-duality lattice identity is exact, so the issues are localized to the BKT section, but that section is load-bearing for the paper's main advertised result.","major_comments":[{"comment":"The claimed spin-wave exponent eta = b^2 is not a consequence of the sampler as defined in Eq. (3.3). The field theta_sw contains an unspecified amplitude A and an implicit mode cutoff, and the distribution p(n) proportional to |n|^{-2} is not normalizable on Z^2 without a regularization. For the Gaussian field (3.3), the vertex correlator is exp[-(b^2/2) Var(theta_sw(x)-theta_sw(0))], and with the stated spectral density the variance behaves as A^2 log|x| / log(Lambda) at large separation. The resulting exponent is b^2 A^2 / (2 log Lambda), not b^2. Equation (3.5) therefore secretly imposes A = sqrt(2 log Lambda) or an equivalent matching condition on the spectral density, which is never stated. This matters because Eq. (3.6) derives b_c = 1/2, K0 = 1/(2 pi b^2), and T_c = pi/2 from eta = b^2; without the hidden normalization these values are inputs, not outputs of the neural sampler.","section":"Section 3, Eq. (3.5)"},{"comment":"The factorization G2(r) = G_sw(r) G_v(r) assumes that the spin-wave and vortex sectors are statistically independent and that the vortex probability in Eq. (3.8) needs no cross-correction from the spin-wave field. In the standard BKT derivation, the logarithmic vortex interaction is obtained after integrating out spin waves, and the vortex fugacity enters with a definite relation to the spin-wave action. The paper does not derive P_vort from the mixed ensemble; it simply writes down the standard Coulomb gas weight with K0 chosen as in Eq. (3.6). As a result, the BKT phenomenology is to a large extent put in by hand rather than produced by the NN-FT construction.","section":"Section 3, Eqs. (3.7)-(3.9)"},{"comment":"The essential singularity claim in Eq. (3.10) involves an unspecified fitted constant c and an unspecified vortex fugacity y, and Figures 1 and 2 show no error bars or sampling parameters such as system size L, mode cutoff, or the parameters A, a_c, and y. Without these details, the statement that the data 'reproduces' the BKT essential singularity and the Nelson-Kosterlitz jump is not quantitatively supported beyond a curve fit with free parameters.","section":"Section 3, Eq. (3.10) and Figures 1-2"}],"minor_comments":[{"comment":"The definition of theta_sw should specify the amplitude A, the mode cutoff Lambda, and the treatment of the n = 0 mode; the distribution p(n) proportional to |n|^{-2} is singular at n = 0 and needs an IR regularization.","section":"Section 3, Eq. (3.3)"},{"comment":"The sentence 'Numerically one finds...' is vague; if the Buscher checks are numerical, the relevant system sizes, sample counts, and tolerances should be reported, and if they are exact, the equations should be stated.","section":"Section 4, paragraph after Eq. (4.8)"},{"comment":"The paper is a proceedings contribution based on [1], but Section 3 should be self-contained enough that a reader does not need the companion paper to understand the origin of, and normalization conditions for, the sampled spin-wave field and vortex fugacity.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The rotor section (Section 5) and the exact T-duality lattice identity (Eq. (4.6)) are solid and give the paper genuine value. The BKT section, however, is the advertised flagship result and its central scaling claim is currently an unstated matching condition rather than a derivation. I would ask the authors to either fix the normalization explicitly and rederive the critical line, or substantially soften the BKT claims and present the section as an illustration of how vortex sectors must be added by hand. This is a fixable issue within the manuscript's scope, hence major_revision rather than reject."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The rotor section works and is the cleanest part of the paper. The BKT spin-wave scaling in Eq. (3.5) is asserted, not derived: the amplitude A and the mode cutoff are never specified, and the claimed η = b^2 only holds once A is tuned to √(2 log Λ). That is a real gap in the central demonstration.\n\nWhat the paper does well: it takes the mixed continuous/discrete construction from companion paper [1] and makes it concrete for two compact bosons. The T-duality section is exact—Eq. (4.6) is the standard lattice sum and the momentum/radius exchange is built in. The rotor in Sec. 5 is a textbook calculation, but the presentation is clear, the MC numbers in Table 1 match the exact correlator, and it makes the point that compactness requires a winding sum on top of the Gaussian sampler.\n\nSoft spots: the BKT section is the load-bearing claim and it has a hidden normalization. For p(n) ∝ |n|^{-2}, the variance of the spin-wave field carries a factor A^2/(2 log Λ), so without fixing A the critical line shifts. The paper never states A or the cutoff, and p(n) isn't normalizable on Z^2 without a regulator. The figures have no error bars and Eq. (3.10) uses a fitted c. These are fixable—state the normalization, or derive it—but as written the BKT 'reproduction' is partly input. The rotor and T-duality parts do not share this flaw.\n\nWho this is for: people at the AI-physics interface who want a template for adding topological sectors to neural samplers. They'll get a useful construction and a clear one-dimensional sanity check. It's a proceedings paper built on [1], so the novelty is modest, but it is honest about that.\n\nRecommendation: send it to a serious referee, but the referee should ask for the missing normalization in Sec. 3. I would not cite the BKT number as a physics result until that is fixed; the rotor check is fine as is.","headline":"The rotor section is clean and exact; the BKT section hides an unstated normalization, so treat the critical line as an illustration rather than a derivation.","tokens_in":8140,"tokens_out":3879,"would_cite":false,"duration_ms":36805,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes that compactness in neural network field theory is a property of the field space rather than of the continuous sampler: the same local Gaussian sector paired with different discrete topological labels gives…","keywords":["neural network field theory","compact boson","BKT transition","T-duality","topological sectors","random Fourier features","Coulomb gas","quantum rotor"],"falsifier":"Run the spin-wave-only sampler of Eq. (3.3) at several amplitudes $A$ and charges $b$ (for example 0.3, 0.5, 0.7) and extract the exponent of $\\langle V_b(x)V_b^*(0)\\rangle$; if the exponent deviates from $b^2$ unless $A$ is tuned, the claimed $\\eta=b^2$ line fails. Independently, in the full mixed ensemble at $b$ slightly below $1/2$, the vortex density should be essentially zero and the pair correlation $g_{+-}(r)$ should stay bound, while just above $1/2$ it should rise; a crossing at a markedly different $b$ would contradict the paper's identification of $b_c$.","tokens_in":6950,"feed_emoji":"🌀","tokens_out":12397,"duration_ms":115781,"temperature":0.7,"pith_summary":"The paper's central claim is that a compact boson in neural network field theory (NN-FT) is not defined by a Gaussian sampler alone: the theory also requires a sum over discrete topological sectors, such as vortices or momentum--winding labels. The paper argues that the same local neural sampler, paired with different discrete topological data, yields physically distinct compact theories, so topology is part of the definition of the field theory rather than an optional addition. It demonstrates this with two case studies: a random Fourier feature spin-wave sampler with an explicit Coulomb gas vortex sector reproduces the qualitative BKT phenomenology, and oscillator modes with momentum--winding labels reproduce circle T-duality, Buscher transformations, self-dual current algebra, and a toy T-fold. A minimal quantum rotor example shows that the mixed continuous/discrete ensemble is exactly equivalent to the conventional Hamiltonian formulation.","feed_headline":"Compact bosons force neural field theory to add discrete sectors","feed_subtitle":"Vortex and winding labels, added to one Gaussian sampler, reproduce BKT physics and T-duality.","key_machinery":"The central object is the mixed latent-variable ensemble $\\Theta=(\\theta,Q)$ of Eq. (1.1). The continuous part $\\theta$ is a random Fourier feature field with Gaussian amplitudes, random phases, and frequencies sampled from a spectral density $\\rho(w)$; at large width it becomes Gaussian with kernel (2.4). The discrete part $Q$ is the sector label: for the BKT boson it is a neutral Coulomb gas configuration with charges $m_a=\\pm1$ and probability (3.8), sampled independently from the spin-wave field so that the two-point function factorizes, $G_2=G_{\\rm sw}G_v$; for the string it is the momentum--winding pair $(n,w)$ with lattice weights (4.6). This object carries the argument because the same continuous sampler, with different choices of $Q$, produces the different physics of the two compact bosons.","core_discovery":"The paper claims that compactness in NN-FT is a property of field space, not of the network architecture: a Gaussian covering field is real and single valued, while a compact boson lives on $S^1$ and carries winding, momentum, vortices, and defects. Consequently the correct ensemble is mixed, $\\Theta=(\\theta,Q)$, with continuous neural parameters $\\theta$ for the local Gaussian sector and discrete labels $Q$ for topological sectors, and observables are computed as in Eq. (1.2). With this ensemble the paper reproduces the BKT transition (Gaussian critical line $\\eta=b^2$ below $T_c$, vortex proliferation above, the essential singularity $\\xi(b)\\sim\\exp(c/\\sqrt{b^2-b_c^2})$, and the Nelson--Kosterlitz jump of the helicity modulus) and the T-duality of the bosonic string (circle duality, Buscher rules on toroidal backgrounds, self-dual $SU(2)_L\\times SU(2)_R$ current algebra enhancement, and a toy T-fold). The same construction yields the exact thermal correlator of the quantum rotor.","pith_inferences":["A natural extension is to apply the same mixed ensemble to lattice gauge theories, promoting flux or charge sectors to discrete latent variables and asking whether confinement/deconfinement transitions emerge from the sampler alone.","The factorization $G_2=G_{\\rm sw}G_v$ suggests that all radius and temperature dependence of the compact theory sits in the discrete sector; if so, topological contributions could be computed without re-simulating the local Gaussian field.","The exact rotor equivalence provides a clean benchmark for finite-width and finite-$N$ corrections in neural-network quantum mechanics, because exact thermal correlators are known for all $\\beta$ and inertia $I$.","If circle T-duality is realized sample-by-sample, then composing Buscher maps along cycles---T-fold transition functions---might be implementable as changes of discrete labels in a patchwise neural sampler, giving a constructive route to non-geometric string backgrounds."],"forward_implications":["Including an explicit sum over discrete topological sectors becomes part of the definition of a compact NN-FT, not an optional addition.","With the vortex sector included, the same random Fourier feature spin-wave sampler reproduces the Gaussian critical line below $T_c$, vortex proliferation above it, the essential singularity of the correlation length, and the Nelson--Kosterlitz jump.","With momentum--winding labels, the same oscillator sampler reproduces circle T-duality, the Buscher rules on toroidal backgrounds, self-dual $SU(2)_L\\times SU(2)_R$ current algebra enhancement, and a toy T-fold.","For the quantum rotor, the mixed continuous/discrete ensemble is exactly equivalent to the conventional Hamiltonian formulation, so the construction is exact in this minimal case.","Without the discrete sector, the compact theory would remain on the Gaussian critical line for all $b$; the topological sum is therefore a dynamical ingredient rather than bookkeeping."],"supporting_citations":[{"why":"The longer paper behind these proceedings, where the full mixed continuous/discrete construction and its derivations appear.","marker":"[1]"},{"why":"Introduces the neural network field theory framework of parameter-space averages over network weights.","marker":"[2]"},{"why":"Provides the generating-functional formulation and the criterion for obtaining Euclidean QFTs from parameter-space averages.","marker":"[3]"},{"why":"Berezinskii's original analysis of destruction of long-range order in two-dimensional systems, the basis of the BKT transition.","marker":"[10]"},{"why":"Kosterlitz and Thouless's ordering/metastability analysis identifying vortex unbinding as the transition mechanism.","marker":"[11]"},{"why":"Kosterlitz's critical properties of the 2D XY model, source of the essential singularity quoted in Eq. (3.10).","marker":"[12]"},{"why":"Buscher's derivation of the symmetry transformation of string background fields used for the Buscher map.","marker":"[14]"},{"why":"Path-integral derivation of quantum duality in nonlinear sigma models, the basis for the toroidal Buscher rules.","marker":"[15]"},{"why":"Random Fourier features, the architecture whose spectral density kernel appears in Eq. (2.4).","marker":"[18]"},{"why":"Nelson and Kosterlitz's universal jump in superfluid density, the target of the helicity modulus comparison.","marker":"[20]"}],"fun_headline_variants":["NN field theory needs discrete sectors for compact bosons","Compact bosons force neural samplers to add topology","One neural sampler, two compact tales: BKT and T-duality","Mixed NN ensemble with vortices and winding cracks compact bosons","Neural field theory gains BKT and T-duality via discrete labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The BKT demonstration depends on an unproven claim about the spin-wave sampler: that with frequencies drawn from $p(n)\\propto |n|^{-2}$, the spin-field two-point function falls as a power set purely by $b$, with no fine-tuning of the amplitude $A$; if that power law shifts, the claimed critical line $\\eta=b^2$ and $b_c=1/2$ move.","fun_headline_variants_meta":{"raw":{"variants":["NN field theory needs discrete sectors for compact bosons","Compact bosons force neural samplers to add topology","One neural sampler, two compact tales: BKT and T-duality","Mixed NN ensemble with vortices and winding cracks compact bosons","Neural field theory gains BKT and T-duality via discrete labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1296,"prompt_tokens":959,"completion_tokens":337,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":250}},"tokens_in":575,"tokens_out":337,"duration_ms":4382,"temperature":1.0,"reasoning_tokens":250,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:03:50.493047+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the spin-wave-only sampler of Eq. (3.3) at several amplitudes $A$ and charges $b$ (for example 0.3, 0.5, 0.7) and extract the exponent of $\\langle V_b(x)V_b^*(0)\\rangle$; if the exponent deviates from $b^2$ unless $A$ is tuned, the claimed $\\eta=b^2$ line fails. Independently, in the full mixed ensemble at $b$ slightly below $1/2$, the vortex density should be essentially zero and the pair correlation $g_{+-}(r)$ should stay bound, while just above $1/2$ it should rise; a crossing at a markedly different $b$ would contradict the paper's identification of $b_c$.","supporting_citations":[{"cited_title":"Nelson and J.M","cited_arxiv_id":null,"evidence_quote":"Nelson and Kosterlitz's universal jump in superfluid density, the target of the helicity modulus comparison."}],"review_version":1}